It is well known that traditional Hidden Markov Models (HMM) systems lead to a considerable improvement when more training data or more parameters are used. However, using more data with hybrid Hidden Markov Models and Artificial Neural Networks (HMM/ANN) models results in increased training times without improvements in performance. We developed in this work a new method based on automatically separating data into several sets and training several neural networks of Multi-Layer Perceptrons (MLP) type on each set. During the recognition phase, models are combined using several criteria (based on data fusion techniques) to provide the recognized word. We showed in this paper that this method significantly improved the recognition accuracy. This method was applied in an Arabic speech recognition system. This last is based on the one hand, on a fuzzy clustering (application of the fuzzy c-means algorithm) and of another share, on a segmentation at base of the genetic algorithms.